Kuznetsova AlexandraStatistics and Data analysis Section, DTU compute

ConsumerCheck: PCA and preference

mapping

1

2

Overview

• ConsumerCheck software

OverviewData sets for PCA and preference mappingPCAPreference mapping Example (apple data)

3

ConsumerCheck

• Standalone software dedicated for analysis of consumer data

• PanelCheck-like software easy-to-use Flexible dedicated for sensory practitioners

• Visualize and analyze your data fast and efficient!

• Classical and advanced statistical methods: Basic statistics (plots/tables) PCA Preference mapping (PLSR, PCR) Conjoint analysis (mixed effects models)

4

ConsumerCheck GUI

Import data Basic stat liking PCA Prefmap PLSR/PCR

5

ConsumerCheck

Example: Apple data5 apples108 consumers

Assessors scoredthe same 5 apples according to 14 attributes

Consumer liking data

Sensory profiling data

6

PCA / Prefmap data sets

Sensory Profile data

Consumer liking data

Product

Consumers

Products Attributes

Data need to be balanced:ALL assessors need to have tested ALLproductsALL consumers need to have tested ALLproducts

7

PCA - Recap

18.08.2015 Oliver Tomic | Nofima | Norway 7

J

K

X

Average consensus

PCA

Scores plot Loadings plot

8

Data Systematic variation + Noise

X = ETP’ +

X: data matrixT: PCA scoresP: PCA loadingsE: residuals /noise

Principal Components (PC’s) describingthe systematic variation in the data

9

ConsumerCheck PCA GUI

Tree control for the plots

By checking the Standartize checkboxAll variables in the data are standardisedSuch that they have zero mean and a standard deviationEquals to one. If unchecked, then only mean centered

Variables with zero variance Across objects (rows)Are left out

How many PC to compute?

10

PCA (sensory profiling data)– Overview plot

PCA ScoresLoadingsCorrelation loadingsExplained variance

11

PCA (sensory profiling data) – Scores

Visualizes how objects (products)Are distributed across spaceSpanned by two principal components (PC)

Product 3 and 5 are very similarProducts 1,2,4 are very different from 3 and 5

PC1 and PC2 describe 99% of the variation in the data

12

PCA (sensory profiling data) – loadings

Visualizes how the variables (sensory attributes) contribute to the variation in the data

Attribute 4 contributes much to the variation explained by PC1

PC2: attributes 5,10 versus 1,2,3,12

13

PCA – correlation loadings

Provides information on how systematic the variance of a variable is with regard to the computed PC’s

Correlation loading = correlation between the original data of a specific variable and the scores of a specific PC

The two rings in the correlation loadings plot indicate amounts of explained variance for the attribute at hand

The outer ring represents 100% of the variance The ionnder 50%

14

PCA – explained variances

Visualizes calibrated and vvalidated explained variances

The validated explained variance is computed by systematically leaving outobjects/rows from the data, then computing new PCA models and using the new loadingsto predict values of the data that were left out. The closer the predictions of the left outdata are to the real values of the left out data, the more robust the model.

Preference mapping

Preference mapping (Greenho and MacFie 1994; McEwan 1996) is a much used statisticalmethod in the field of sensometrics that analyses consumer liking and dsensory profling data together.

15

16

Preference mapping GUI

Row order of the products needs to be the same

Preference mapping

17

Consumer data

PCA of consumer data

Sensory loadings

Relate each sensory attributeto these directions

Sensory data

PCA of sensory data

Consumer loadings

Relate each consumerto these directions

X X

18

Preference mapping – X scores

Here: internal preference mappingPLSRX and Y not standardized (all variables are based on the same scale)

X-scores: visualizes how the products related to each other in the space spanned by the first principal components

Only 68% from consumer liking data

19

Preference mapping – X and Y correlation loadings

Correlation loadings from both X and Y are visualized

Correlation loadings belonging to X are blue

Many consumers prefer Products with high intensities ofAttribute 2 and 6

20

Preference mapping – X loadings

X loadings: show how the variables of the X matrix (here consumer data) contribute to the common variation between X and Y for each PC

21

Preference mapping – Y correlation loadings

Y loadings: show how the variables of the Y matrix contribute to the common variation between X and Y for each PC

22

Preference mapping – Y correlation loadings

Y loadings: show how the variables of the Y matrix contribute to the common variation between X and Y for each PC

23

Summary ConsumerCheck

• Easy-to-use software for non-statisticians

• Proposes classical as well as advanced tools for analysis of consumer data

• Important methods:

+ PCA+ Preference mapping+ Conjoint analysis